Hyperspectral detection method for uniformly mixed material components of sintered ore
By combining multi-scale attention convolutional neural networks and wavelet transform, the problems of time lag and insufficient accuracy in the detection of sintered ore blending materials were solved, and real-time, high-precision online analysis of hyperspectral detection was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-18
- Publication Date
- 2026-03-27
AI Technical Summary
Existing technologies for detecting the composition of sintered ore blends suffer from time lag, insufficient detection accuracy, and poor adaptability, making it impossible to achieve real-time, high-precision online detection.
A hyperspectral detection method combining multi-scale attention convolutional neural networks and wavelet transform is adopted. By constructing a multi-scale attention convolutional neural network model, data collected by a hyperspectral camera is preprocessed and trained to establish a mapping relationship between component content, thereby achieving high-precision component prediction.
It improves detection sensitivity and accuracy, enables non-contact, large-area, rapid scanning, adapts to material pile inhomogeneity, and meets the requirements for real-time, high-precision online detection.
Smart Images
Figure CN121746920A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to sinter mixing material composition analysis, in particular to a sinter mixing material composition hyperspectral detection method based on wavelet transform and multi-scale attention convolutional neural network, which is suitable for sinter mixing material composition detection in ironworks and belongs to the technical field of spectral analysis. BACKGROUND
[0002] As a key link of raw material pretreatment in steel production, the composition stability (such as uniform distribution of total iron content TFe, silicon dioxide SiO2, aluminum oxide Al2O3 and other contents) of sinter mixing material composition analysis is directly related to the stability of sinter quality, the smooth running of blast furnace and the control of energy consumption. The traditional mixing material composition detection mainly relies on manual sampling and laboratory analysis, which has significant time lag (usually several hours are needed), and cannot meet the real-time regulation and control requirements of modern steel industry on raw material preparation. This lag leads to the delay of mixing process adjustment, which not only affects the uniformity of sinter quality, but also may cause the fluctuation of blast furnace operation, resulting in energy waste and increase of production cost.
[0003] Current methods for analyzing the composition of homogenized materials primarily employ offline methods such as manual sampling combined with X-ray fluorescence spectroscopy (XRF) or chemical analysis. These methods involve multiple steps, including sampling, drying, and sample preparation, resulting in a lengthy process and hindering continuous, high-frequency analysis. Furthermore, manual sampling suffers from insufficient spatial representativeness, susceptibility to sample contamination, and significant operational errors, making it difficult to accurately reflect the overall compositional distribution of the material pile and providing timely and effective guidance for real-time material distribution and extraction. In recent years, online detection technologies based on spectral analysis have been increasingly applied to raw material composition analysis. This technology identifies the spectral characteristics of materials, establishes a quantitative relationship model between spectral information and chemical composition, and achieves rapid composition inversion. Commonly used spectroscopic techniques in solid raw material analysis include near-infrared spectroscopy (NIRS), laser-induced breakdown spectroscopy (LIBS), and X-ray fluorescence spectroscopy. Near-infrared spectroscopy is widely used for detecting indicators such as moisture content and organic matter due to its portability and ease of operation. However, it has low sensitivity to metal oxide components and is easily affected by factors such as particle size and moisture content. Laser-induced breakdown spectroscopy has the ability to detect multiple elements simultaneously, but when faced with uneven particle size and large variations in surface roughness in mixed materials, the plasma excitation stability is poor, and the matrix effect is significant, affecting the quantitative accuracy. Existing component analysis models mostly rely on traditional machine learning methods such as partial least squares (PLS) and support vector machines (SVM). These methods have limited feature extraction capabilities under nonlinear effects such as complex particle scattering and moisture interference, and are difficult to adapt to the spatial heterogeneity of component distribution and operating condition fluctuations in mixed material piles. Summary of the Invention
[0004] To address the aforementioned shortcomings of existing technologies, the present invention aims to propose a hyperspectral analysis method for sintered ore blends. This invention exhibits strong adaptability to the non-uniformity of the sinter pile, significantly improves detection sensitivity and accuracy, and enables online, real-time, high-precision detection of the components of sintered ore blends.
[0005] The technical solution of this invention is implemented as follows:
[0006] A hyperspectral method for detecting the composition of sintered ore blends, comprising the following steps:
[0007] 1) Obtain different sintered ore blend samples. All sintered ore blend samples contain all the components to be detected and the content of each component is known. The content of any one component to be detected is different in different batches. Use a hyperspectral camera to collect hyperspectral data of all sintered ore blend samples. Divide the hyperspectral data into training set and test set.
[0008] 2) Construct a multi-scale attention convolutional neural network model, which includes an input layer, multiple convolutional blocks, a global average pooling layer, and a fully connected layer; each convolutional block contains a convolutional layer, a batch normalization layer, a ReLU activation function, a multi-scale feature extraction module, and a max pooling layer connected in sequence.
[0009] 3) Input the hyperspectral data corresponding to the training set into the multi-scale attention convolutional neural network model. Continuously train the multi-scale attention convolutional neural network model with the hyperspectral data corresponding to the training set to establish the mapping relationship between the hyperspectral data and the content of the component to be detected. Validate the model performance with the test set. The trained multi-scale attention convolutional neural network model is the component prediction model.
[0010] 4) Use a hyperspectral camera to collect hyperspectral data of the sintered ore blend sample to be tested, input the hyperspectral data of the sintered ore blend sample to be tested into the composition prediction model, and the composition prediction model outputs the predicted results of the components to be detected and their contents in the sintered ore blend.
[0011] Further, in step 1), the hyperspectral data of each sintered ore blend sample is preprocessed. The preprocessing includes background noise subtraction, wavelet transform, and data normalization performed sequentially. The wavelet transform uses a multi-level discrete wavelet transform to decompose each spectral signal, obtaining the corresponding approximation coefficients and detail coefficients. The mathematical formula is as follows:
[0012] in, The Jth order approximation coefficient, For the j-th level detail coefficients, the approximation coefficient length and detail coefficient length are reconstructed to the original signal length using linear interpolation, and finally the original signal, approximation coefficients and detail coefficients are stacked to form multiple channel features.
[0013] Furthermore, the background noise subtraction is performed by subtracting background noise from the collected sample hyperspectral data using hyperspectral data without samples.
[0014] Furthermore, in step 2), the number of input channels in the input layer of the multi-scale attention convolutional neural network model is equal to the number of multi-channel features constructed by wavelet transform.
[0015] Furthermore, the multi-scale feature extraction module uses three convolutional kernels of different scales to input hyperspectral data in parallel, then concatenates the outputs of the three branches along the channel dimension, and then restores the number of channels to the number of input channels using a unit convolutional kernel to obtain the fused feature map. The fused feature map is then subjected to a channel attention mechanism to calculate and weight the importance weight of each channel, and then a spatial attention mechanism is applied to the channel-weighted feature map to calculate and weight the importance weight of each spatial location to obtain the output feature map.
[0016] Furthermore, in step 4), the hyperspectral data of the collected sintered ore mixture to be tested undergoes the same preprocessing as the hyperspectral data of the sintered ore mixture sample in step 1).
[0017] Furthermore, in step 3), the mean squared error is used as the loss function during the training of the multi-scale attention convolutional neural network model:
[0018] in, This represents the actual concentration value. To predict concentration values.
[0019] Furthermore, in step 3), the Adam optimizer is used to optimize and update all parameters in the model, and the loss function is minimized through continuous iterative training; at the same time, the learning rate scheduler is used to dynamically adjust the learning rate, and when the validation set loss no longer decreases, the learning rate is reduced by a certain proportion to improve the generalization ability of the model.
[0020] Compared with the prior art, the present invention has the following beneficial effects:
[0021] 1. The multi-scale feature extraction module of this invention combines multi-scale convolution and a dual attention mechanism. By using convolution kernels of different sizes, it can simultaneously capture local details and global trends, achieving efficient extraction of multi-scale features from spectral data. Through the dual attention mechanism, the model can adaptively select important features and suppress unimportant features, cascading enhancements ultimately improving feature quality and significantly enhancing detection sensitivity and accuracy.
[0022] 2. This invention uses wavelet transform to preprocess the original hyperspectral signal, which can simultaneously capture the global trend (low frequency) and local details (high frequency) of the signal. It can also reduce the impact of noise by emphasizing or suppressing certain frequency bands, which helps the model understand spectral features at different scales. Furthermore, by using wavelet coefficients as additional channels, the model can be provided with richer feature representations, further improving detection accuracy.
[0023] 3. This invention eliminates the need for complex sample preparation and enables non-contact, large-area rapid scanning, making it more suitable for online monitoring needs in harsh environments of sintering raw material fields. It provides a more promising solution for real-time, high-precision analysis of the composition of mixed materials. Attached Figure Description
[0024] Figure 1 This is a flowchart of the hyperspectral detection process for the composition of the sintered ore blending material of the present invention.
[0025] Figure 2 This is an architecture diagram of the multi-scale feature extraction module of the present invention.
[0026] Figure 3 This is a diagram of the architecture of the multi-scale attention convolutional neural network module of this invention. Detailed Implementation
[0027] Hyperspectral analysis, as a novel detection method integrating imaging and spectral analysis, continuously acquires hundreds of narrow-band spectral images, allowing each pixel to obtain a complete spectral curve, thus demonstrating unique advantages in the spatial distribution analysis of homogenized materials. Compared with traditional spectral analysis techniques, hyperspectral technology has three major advantages: 1. Its high spectral resolution can accurately identify the weak characteristic absorption peaks of different mineral phases in the homogenized material, significantly improving the detection sensitivity and accuracy of major components such as TFe and SiO2; 2. Combining spatial and spectral dimensions, deep learning algorithms can automatically learn the correlation between morphological features such as particle size and distribution uniformity and composition, enhancing the model's adaptability to material inhomogeneity; 3. This technology requires no complex sample preparation, enabling non-contact, large-area rapid scanning, making it more suitable for the online monitoring needs in the harsh environment of sintering raw material fields, providing a more promising solution for real-time high-precision analysis of homogenized material composition.
[0028] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0029] See Figure 1 This invention discloses a hyperspectral method for detecting the composition of sintered ore blends, comprising the following steps:
[0030] 1) Obtain multiple batches of sintered ore blend samples. All sintered ore blend samples contain all the components to be detected and the content of each component is known. The content of any one component to be detected is different in different batches. Use a hyperspectral camera to collect hyperspectral data of all sintered ore blend samples. Perform background noise subtraction, wavelet transform and data standardization on the collected hyperspectral data in sequence. Divide the preprocessed data into a training set and a test set.
[0031] 2) Construct a multi-scale attention convolutional neural network (CNN) model, which includes an input layer, four convolutional blocks, a global average pooling layer, and two fully connected layers. Each convolutional block contains a convolutional layer, a batch normalization layer, a ReLU activation function, a multi-scale attention block (MAB), and a max pooling layer connected in sequence. The number of input channels in the input layer is equal to the number of multi-channel features constructed by wavelet transform.
[0032] 3) The hyperspectral data corresponding to the training set is input into the multi-scale attention convolutional neural network model. The model is continuously trained using the hyperspectral data from the training set to establish a mapping relationship between the hyperspectral data and the content of the components to be detected. The model performance is then validated and optimized using test set data. The trained multi-scale attention convolutional neural network model is the component prediction model. The model input consists of multi-channel features constructed after wavelet transform. After feature extraction through four convolutional blocks, global average pooling is used to compress the features of each channel into a scalar, which is then output as seven regression prediction values through two fully connected layers.
[0033] 4) Use a hyperspectral camera to acquire hyperspectral data of the sintered ore blend sample to be tested. Input the hyperspectral data of the sintered ore blend sample to be tested into the composition prediction model. The composition prediction model will then output the predicted results of the components to be detected and their contents in the sintered ore blend. It should be noted that the acquired hyperspectral data of the sintered ore blend to be tested undergoes the same preprocessing as the hyperspectral data of the sintered ore blend sample in step 1) before being input into the composition prediction model.
[0034] Step 1) Preprocessing of hyperspectral data: To eliminate the influence of ambient light and other possible factors, the background noise subtraction method uses hyperspectral data without samples to subtract background noise from the collected sample hyperspectral data. The hyperspectral data after background subtraction is then subjected to wavelet transform. Wavelet transform is a signal processing technique that decomposes a signal into subbands of different frequencies, thereby providing a time-frequency representation of the signal. The wavelet transform method uses multi-level discrete wavelet transform to decompose each spectral signal, obtaining corresponding approximation coefficients and detail coefficients. The mathematical formula is as follows:
[0035]
[0036] in The Jth order approximation coefficient, For the j-th level detail coefficients, since the wavelet coefficient length is not equal to the original signal length, linear interpolation is used to reconstruct the original length. Finally, the original signal, approximation coefficients, and detail coefficients are stacked to form multiple channel features. These multi-channel features are used as input data and undergo data standardization processing. The data standardization processing is the standard deviation standardization (Z-score) method, the core formula of which is:
[0037]
[0038] in The standardized value. These are the original eigenvalues. The mean of the features, The standard deviation is the characteristic.
[0039] See Figure 3 The multi-scale attention convolutional neural network model constructed in step 2) of this invention comprises an input layer, four convolutional blocks (each containing a convolutional layer, a batch normalization layer, a ReLU activation function, a multi-scale feature extraction module, and a max pooling layer), a global average pooling layer, and two fully connected layers. The core of the entire network is the multi-scale feature extraction module, see [link to relevant documentation]. Figure 2 This module uses three convolutional kernels of different sizes to input hyperspectral data feature maps in parallel. The outputs of the three branches are then concatenated along the channel dimension, and a single convolutional kernel is used to restore the channel number to the corresponding input channel number, resulting in a fused feature map. A channel attention mechanism is then applied to the fused feature map, calculating and weighting the importance weight of each channel. A spatial attention mechanism is then applied to the channel-weighted feature map, calculating and weighting the importance weight of each spatial location, resulting in an output feature map. This invention, by using convolutional kernels of different sizes, can simultaneously capture local details and global trends. This helps the model better understand different parts of the spectrum; by concatenating features at different scales, richer feature representations can be formed; through the dual attention mechanism, the model can adaptively select important features and suppress unimportant features; and through multiple convolutional layers and max-pooling layers, the multi-scale attention convolutional neural network can learn hierarchical representations from low-level features to high-level features.
[0040] In step 3) of this invention, during the model training phase, the mean squared error (MSE) is used as the loss function to measure the difference between the predicted value and the true label, as shown in the following formula:
[0041] in, This represents the actual concentration value. To predict concentration values.
[0042] The model training utilizes the Adam optimizer, the ReduceLROnPlateau learning rate scheduler, and Dropout regularization. The Adam optimizer optimizes and updates all parameters in the model, iteratively training to minimize the loss function, allowing the model to gradually learn the complex mapping relationship between spectral data and substance content. Simultaneously, the learning rate scheduler dynamically adjusts the learning rate, reducing it proportionally when the validation set loss no longer decreases, thus improving the model's generalization ability. Furthermore, to prevent overfitting, Dropout is used to randomly mask neurons. In this example, the Adam optimizer is set with an initial learning rate lr = 0.001, a weight decay coefficient weight_decay = 1e-5, and other parameters β1 = 0.9, β2 = 0.999, ε = 1e-8. The ReduceLROnPlateau learning rate scheduler is set with a learning rate decay factor factor = 0.5 and a patience value patience = 50. Dropout is configured to randomly drop neurons with probability p = 0.5 during the training phase, and activate all neurons during the testing phase, multiplying the output by (1-p).
[0043] As can be seen from the above introduction, this invention has made improvements and innovations in the following aspects to improve the accuracy of the model in detecting the composition of sintered ore blends:
[0044] 1. In terms of data preprocessing, to improve the detection accuracy of subsequent models, background noise is removed from the acquired hyperspectral data using dark spectroscopy. Wavelet transform is used to decompose the original hyperspectral signal into approximate coefficients (low frequency) and detail coefficients (high frequency). These coefficients are then adjusted to the original signal length using interpolation methods and input as multi-channel features into the convolutional neural network. This processing can simultaneously capture the global trend (low frequency) and local details (high frequency) of the signal, and can also reduce the impact of noise by emphasizing or suppressing certain frequency bands, which helps the model understand spectral features at different scales. In addition, by using wavelet coefficients as additional channels, richer feature representations can be provided to the model. Z-score normalization is used to eliminate the influence of dimensions and numerical range, and to accelerate model convergence and optimize training stability. The above preprocessing steps effectively eliminate spectral noise and enhance feature representation.
[0045] 2. Feature extraction of hyperspectral signals using a multi-scale attention convolutional neural network. The convolutional neural network constructed in this invention comprises four convolutional blocks, one global average pooling layer, and two fully connected layers. The multi-scale feature extraction module in each convolutional block, by using convolutional kernels of different sizes, can simultaneously capture local details (such as sharp peaks) and global trends (such as baseline drift). This helps the model better understand different parts of the spectrum; by concatenating features at different scales, richer feature representations can be formed; and through a dual attention mechanism, the model can adaptively select important features and suppress unimportant features. Through multiple convolutional layers and a max-pooling layer, the multi-scale attention convolutional neural network can learn hierarchical representations from low-level features (such as edges and peaks) to high-level features (such as the entire spectral pattern), which facilitates more complex inference by the model.
[0046] The weighted fusion features are input into the prediction network, and the fusion features are further nonlinearly transformed and combined through fully connected layers and ReLU activation functions, finally outputting the predicted values of the content of multiple components of the detected sample.
[0047] To verify the effectiveness of this application, the inventors conducted the following verification experiments:
[0048] First, hyperspectral data of 20 different batches of molten iron blocks with known main component contents (Si, S) were collected using a hyperspectral camera. Ten sampling areas were selected on each molten iron block sample, and the acquisition scheme was repeated once a day for a total of six days. The wavelength range of the camera sampling was 391.87 nm to 1003.87 nm, and the sampling interval was approximately 2.1 nm.
[0049] The hardware environment for this experiment was: a 12th Gen Intel(R) Core(TM) i7-12650H (2.30 GHz) processor and an NVIDIA RTX 4060 graphics card; the software environment was: FigSpec system, Python 3.12.7 interpreter, PyTorch 2.6.0 deep learning framework; and CUDA version 11.8.
[0050] In this experiment, the preprocessing part sets the wavelet basis of wavelet transform to Dobermann wavelet, the filter length N=4, the decomposition level level=3, the original signal length to 300, A1=150, D1=150, A2=75, D2=75, A3=38, D3=38.
[0051] For different batches of sintered ore blend samples, two verification schemes were set up: The first scheme was to use 20 sets of sample data collected on the same day, with the first 1-16 sets of sample data as the training set and the last 17-20 sets of data as the test set; The second scheme was to use 20 sets of sample data collected on different days, with the 20 sets of data collected on one day as the training set and the 20 sets of data collected on another day as the test set.
[0052] The three branches of the multi-scale feature extraction module have convolutional kernels of 3×3, 5×5, and 7×7, respectively, with padding=1. Finally, a 1×1 convolutional kernel is used to compress the channels of the three branches into the original number of channels, and then a dual attention mechanism is used to obtain the fused features.
[0053] The number of input channels of the multi-scale attention convolutional neural network is equal to the number of multi-channel features constructed by wavelet transform (a total of 5 channels, corresponding to the original signal, approximation coefficient A3, detail coefficients D3, detail coefficients D2, and detail coefficients D1, respectively). The parameters of the convolutional blocks are 32 channels (kernel_size=5, padding=1), 64 channels (kernel_size=5, padding=1), 128 channels (kernel_size=3, padding=1), and 256 channels (kernel_size=3, padding=1), respectively. Max pooling (kernel_size=2, stride=2) is used after each convolutional block, and then global average pooling is used to compress the feature map into a 1-dimensional vector. Finally, 7 vectors are output through two fully connected layers.
[0054] Configure the model's optimizer as Adam, training epochs as 400, batch size as 8, initial learning rate as 0.001, weight decay as 1e-5, and learning rate scheduler as ReduceLROnPlateau with patience as 50 and factor as 0.5. Record the final component prediction performance of the model.
[0055] To evaluate the predictive performance of the model, the repeatability and accuracy of the model's predictions of several specific components (TFe, SiO2, CaO, MgO, Al2O3, TiO2) in the sintered ore blend were evaluated, referencing the internal standards of a large steel company. The results are shown in Tables 1 and 2. Table 1 shows the measurement results of the verification experiment for Scheme 1; Table 2 shows the measurement results of the verification experiment for Scheme 2.
[0056] Table 1. Results of the verification experiment for Scheme 1
[0057] Component name Minimum absolute error (%) Maximum absolute error (%) Average absolute error (%) Static accuracy requirement (%) TFe 0.240 0.360 0.280 ≤0.35 SiO2 0.010 0.230 0.120 ≤0.27 Al2O3 0.110 0.640 0.340 ≤0.39 CaO 0.050 0.110 0.090 ≤0.22 MgO 0.040 0.230 0.100 ≤0.39 TiO2 0.030 0.010 0.020 ≤0.04
[0058] Table 2. Results of the verification experiment for scheme two.
[0059] Component name Minimum absolute error (%) Maximum absolute error (%) Average absolute error (%) Static accuracy requirement (%) TFe 0.002 1.049 0.393 ≤0.35 SiO2 0.001 0.472 0.135 ≤0.27 Al2O3 0.001 1.062 0.314 ≤0.39 CaO 0.001 0.577 0.114 ≤0.22 MgO 0.0002 0.289 0.090 ≤0.39 TiO2 0.0001 0.066 0.020 ≤0.04
[0060] The results of the verification experiments in Tables 1 and 2 show that the measurement accuracy for the target components meets the static accuracy requirements, indicating that the measurement accuracy of the present invention meets the standards for practical industrial applications.
[0061] Finally, it should be noted that the specific examples mentioned above are only for explaining the present invention and do not constitute a limitation on the embodiments of the present invention. Although the present invention has selected preferred examples and described them accordingly, those skilled in the art can make other forms of adjustments and improvements based on the above explanations. It is impossible to list all possible embodiments here. Any adjustments and improvements that fall directly or indirectly within the scope of the technical solutions of the present invention can be considered within the protection scope of the present invention.
Claims
1. A hyperspectral method for detecting the composition of sintered ore blends, characterized in that: The steps are as follows: 1) Obtain different sintered ore blend samples. All sintered ore blend samples contain all the components to be detected and the content of each component is known. The content of any one component to be detected is different in different batches. Use a hyperspectral camera to collect hyperspectral data of all sintered ore blend samples. Divide the hyperspectral data into training set and test set. 2) Construct a multi-scale attention convolutional neural network model, which includes an input layer, multiple convolutional blocks, a global average pooling layer, and a fully connected layer; each convolutional block contains a convolutional layer, a batch normalization layer, a ReLU activation function, a multi-scale feature extraction module, and a max pooling layer connected in sequence. 3) Input the hyperspectral data corresponding to the training set into the multi-scale attention convolutional neural network model. Continuously train the multi-scale attention convolutional neural network model with the hyperspectral data corresponding to the training set to establish the mapping relationship between the hyperspectral data and the content of the component to be detected. Validate the model performance with the test set. The trained multi-scale attention convolutional neural network model is the component prediction model. 4) Use a hyperspectral camera to collect hyperspectral data of the sintered ore blend sample to be tested, input the hyperspectral data of the sintered ore blend sample to be tested into the composition prediction model, and the composition prediction model outputs the predicted results of the components to be detected and their contents in the sintered ore blend.
2. The method for hyperspectral detection of the composition of sintered ore blends according to claim 1, characterized in that: In step 1), the hyperspectral data of each sintered ore blend sample is preprocessed. The preprocessing includes background noise subtraction, wavelet transform, and data normalization performed sequentially. The wavelet transform uses a multi-level discrete wavelet transform to decompose each spectral signal, obtaining the corresponding approximation coefficients and detail coefficients. The mathematical formula is as follows: ； in, The Jth order approximation coefficient, For the j-th level detail coefficients, the approximation coefficient length and detail coefficient length are reconstructed to the original signal length using linear interpolation, and finally the original signal, approximation coefficients and detail coefficients are stacked to form multiple channel features.
3. The method for hyperspectral detection of the composition of sintered ore blends according to claim 2, characterized in that: The background noise subtraction is performed by subtracting background noise from the collected sample hyperspectral data using hyperspectral data without samples.
4. The method for hyperspectral detection of the composition of sintered ore blends according to claim 2, characterized in that: In step 2), the number of input channels in the input layer of the multi-scale attention convolutional neural network model is equal to the number of multi-channel features constructed by wavelet transform.
5. The hyperspectral detection method for the composition of sintered ore blends according to claim 1, characterized in that: The multi-scale feature extraction module uses three convolutional kernels of different scales to input hyperspectral data in parallel, then concatenates the outputs of the three branches in the channel dimension, and then restores the number of channels to the number of channels corresponding to the input channels through a unit convolutional kernel to obtain the fused feature map. The channel attention mechanism is applied to the fused feature map to calculate and weight the importance weight of each channel. Then, the spatial attention mechanism is applied to the channel-weighted feature map to calculate and weight the importance weight of each spatial location, thus obtaining the output feature map.
6. The method for hyperspectral detection of the composition of sintered ore blends according to claim 1, characterized in that: In step 4), the hyperspectral data of the collected sintered ore mixture to be tested are preprocessed in the same way as the hyperspectral data of the sintered ore mixture sample in step 1).
7. The method for hyperspectral detection of the composition of sintered ore blends according to claim 1, characterized in that: In step 3), the mean squared error is used as the loss function during the training of the multi-scale attention convolutional neural network model. ； in, This represents the actual concentration value. To predict concentration values.
8. The method for hyperspectral detection of the composition of sintered ore blends according to claim 1, characterized in that: In step 3), the Adam optimizer is used to optimize and update all parameters in the model. Through iterative training, the loss function is minimized. At the same time, the learning rate scheduler is used to dynamically adjust the learning rate. When the loss on the validation set no longer decreases, the learning rate is reduced by a certain proportion to improve the generalization ability of the model.